It has become increasingly common for Unix-based computer applications to be hosted on a cluster that includes a plurality of computers. It is a goal of cluster operating systems to render operation of the cluster as transparent to applications/users as if it were a single computer. For example, a cluster typically provides a global file system that enables a user to view and access all conventional files on the cluster no matter where the files are hosted. This transparency does not, however, extend to device access on a cluster.
Typically, device access on Unix-based systems is provided through a special file system (e.g., SpecFS) that treats devices as files. This special file system operates only on a single node. That is, it only allows a user of a particular node to view and access devices on that node, which runs counter to the goal of global device visibility on a cluster. These limitations are due to the lack of coordination between the special file systems running on the various nodes as well as a lack of a device naming strategy to accommodate global visibility of devices. These aspects of a prior art device access system are now described with reference to FIGS. 1-4. referring to FIG. 1, there is shown a block diagram of a conventional computer system 100 that includes a central processing unit (CPU) 102, a high speed memory 104, a plurality of physical devices 106 and a group of physical device interfaces 108 (e.g., busses or other electronic interfaces) that enable the CPU 102 to control and exchange data with the memory 102 and the physical devices 106. The memory 102 can be a random access memory (RAM) or a cache memory.
The physical devices 106 can include but are not limited to high availability devices 112, printers 114, kernel memory 116, communications devices 118 and storage devices 120 (e.g., disk drives). Printers 114 and storage devices 120 are well-known. High availability devices 112 include devices such as storage units or printers that have associated secondary devices. Such devices are highly available as the secondary devices can fill in for their respective primary device upon the primary's failure. The kernel memory 116 is a programmed region of the memory 102 that includes accumulating and reporting system performance statistics. The communications devices 118 include modems, ISDN interface cards, network interface cards and other types of communication devices. The devices 106 can also include pseudo devices 122, which are software devices not associated with an actual physical device.
The memory 104 of the computer 100 can store an operating system 130, application programs 150 and data structures 160. The operating system 130 executes in the CPU 102 as long as the computer 100 is operational and provides system services for the processor 102 and applications 150 being executed in the CPU 102. The operating system 130, which is modeled on v. 2.6 of the Solaris.TM. operating system employed on Sun.RTM. workstations, includes a kernel 132, a file system 134, device drivers 140 and a device driver interface (DDI) framework 142. Solaris and Sun are trademarks and registered trademarks, respectively, of Sun Microsystems, Inc. The kernel 116 handles system calls from the applications 150, such as requests to access the memory 104, the file system 134 or the devices 106. The file system 134 and its relationship to the devices 106 and the device drivers 140 is described with reference to FIGS. 2A and 2B.
Referring to FIG. 2A, there is shown a high-level representation of the file system 134 employed by v. 2.6 and previous versions of the Solaris operating system. In Solaris, the file system 134 is the medium by which all files, devices 106 and network interfaces (assuming the computer 100 is networked) are accessed. These three different types of accesses are provided respectively by three components of the file system 134: a Unix file system 138u (UFS), a special file system 138s (SpecFS) and a network file system 138n (NFS).
In Solaris, an application 150 initially accesses a file, device or network interface (all referred to herein as a target) by issuing an open request for the target to the file system 134 via the kernel 132. The file system 134 then relays the request to the UFS 138u, SpecFS 138s or NFS 138n, as appropriate. If the target is successfully opened, the UFS, SpecFS or NFS returns to the file system 134 a vnode object 136 that is mapped to the requested file, device or network node. The file system 134 then maps the vnode object 136 to a file descriptor 174, which is returned to the application 150 via the kernel 132. The requesting application subsequently uses the file descriptor 174 to access the corresponding file, device or network node associated with the returned vnode object 136.
The vnode objects 136 provide a generic set of file system services in accordance with a vnode/VFS interface or layer (VFS) 172 that serves as the interface between the kernel 132 and the file system 134. Solaris also provides inode, snode and mode objects 136i, 136s, 136r that inherit from the vnode objects 136 and also include methods and data structures customized for the types of targets associated with the UFS, SpecFS and NFS, respectively. These classes 136i, 136s and 136r form the low level interfaces between the vnodes 136 and their respective targets. Thus, when the UFS, SpecFS or NFS returns a vnode object, that object is associated with a corresponding inode, snode or mode that performs the actual target operations. Having discussed the general nature of the Solaris file system, the focus of the present discussion will now shift to the file-based device access methods employed by Solaris.
Referring to FIG. 2B, Solaris applications 150 typically issue device access requests to the file system 134 (via the kernel 132) using the logical name 166 of the device they need opened. For example, an application 150 might request access to a SCSI device with the command: open(/dev/dsk/disk.sub.-- logical.sub.-- address). The logical name, /dev/dsk/disk.sub.-- logical.sub.-- address, indicates that the device to be opened is a disk at a particular logical address. In Solaris, the logical address for a SCSI disk might be "c0t0d0sx", where "c0 " represents SCSI controller 0, t0 represents target 0, d0 represents disk 0, and sx represents the xth slice for the particular disk (a SCSI disk drive can have as many as eight slices).
The logical name is assigned by one of the link generators 144, which are user-space extensions of the DDI framework 142, and is based on information supplied by the device's driver 140 upon attachment of the device and a corresponding physical name for the device generated by the DDI framework 142. When an instance of a particular device driver 140 is attached to the node 100, the DDI framework 142 calls the attach routine of that driver 140. The driver 140 then assigns a unique local identifier to and calls the ddi.sub.-- create.sub.-- minor.sub.-- nodes method 146 of the DDI framework 142 for each device that can be associated with that instance. Typically, the unique local identifier constitutes a minor name (e.g., "a") and a minor number (e.g., "2"). Each time it is called, the ddi.sub.-- create.sub.-- minor.sub.-- nodes method 146 creates a leaf node in the DevInfo tree 162 that represents a given device. For example, because a SCSI drive (i.e., instance) can have up to eight slices (i.e., devices), the local SCSI driver 140 assigns unique local identifiers to each of the eight slices and calls the ddi.sub.-- create.sub.-- minor.sub.-- nodes method 146 with the local identifiers up to eight times.
Also associated with each device 106 is a UFS file 170 that provides configuration information for the target device 106. The name of a particular UFS file 170i is the same as a physical name 168i derived from the physical location of the device on the computer. For example, a SCSI device might have the following physical name 168, /devices/iommu/sbus/esp1/sd@addr:minor.sub.-- name, where addr is the address of the device driver sd and minor.sub.-- name is the minor.sub.-- name of the device instance, which is assigned by the device driver sd. How physical names are derived is described below in reference to FIG. 3.
To enable it to open a target device given the target device's logical name, the file system 134 employs a logical name space data structure 164 that maps logical file names 166 to physical file names 168. The physical names of devices 106 are derived from the location of the device in a device information (Devinfo) tree 140 (shown in FIG. 1), which represents the hierarchy of device types, bus connections, controllers, drivers and devices associated with the computer system 100. Each file 170 identified by a physical name 168 includes in its attributes an identifier, or dev.sub.-- t (short for device type), which is uniquely associated with the target device. This dev.sub.-- t value is employed by the file system 134 to access the correct target device via the SpecFS 138s. It is now described with reference to FIG. 3 how dev.sub.-- t values are assigned and the DevInfo tree 140 maintained by the DDI framework 142.
Referring to FIG. 3, there is shown an illustration of a hypothetical DevInfo tree 162 for the computer system 100. Each node of the DevInfo tree 162 corresponds to a physical component of the device system associated with the computer 100. Different levels correspond to different levels of the device hierarchy. Nodes that are directly connected to a higher.sub.-- node represent objects that are instances of the higher level object. Consequently, the root node of the DevInfo tree is always the "/" node, under which the entire device hierarchy resides. The intermediate nodes (i.e., nodes other than the leaf and leaf-parent nodes) are referred to as nexus devices and correspond to intermediate structures, such as controllers, busses and ports. At the next to bottom level of the DevInfo tree are the device drivers, each of which can export, or manage, one or more devices. At the leaf level are the actual devices, each of which can export a number of device instances, depending on the device type. For example, a SCSI device can have up to seven instances.
The hypothetical DevInfo tree 162 shown in FIG. 3 represents a computer system 100 that includes an input/output (i/o) controller for memory mapped i/o devices (iommu) at a physical address addr0. The iommu manages the CPU's interactions with i/o devices connected to a system bus (sbus) at address addr1 and a high speed bus, such as a PCI bus, at address addr2. Two SCSI controllers (esp1 and esp2) at respective addresses addr3 and addr4 are coupled to the sbus along with an asynchronous transfer mode (ATM) controller at address addr5. The first SCSI controller esp1 is associated with a SCSI device driver (sd) at address 0 (represented as @0) that manages four SCSI device instances (dev0, dev1, dev2, dev3). Each of these device instances corresponds to a respective slice of a single, physical device 106. The first SCSI controller esp1 is also associated with a SCSI device driver (sd) at address 1 that manages plural SCSI device instances (not shown) of another physical device 106.
Each type of device driver that can be employed with the computer system 100 is assigned a predetermined, unique major number. For example, the SCSI device driver sd is assigned the major number 32. Each device is associated with a minor number that, within the group of devices managed by a single device driver, is unique. For example, the devices dev0, dev1, dev2 and dev3 associated with the driver sd at address 0 have minor numbers 0, 1, 2 and 3 and minor names a, b, c, d, respectively. Similarly, the devices managed by the driver sd at address 1 would have minor numbers distinct from those associated with the devices dev0-dev3 (e.g., four such might have minor numbers 4-7). The minor numbers and names are assigned by the parent device driver 140 (FIG. 1) for each new device instance (recall that a SCSI instance might be a particular SCSI drive and a SCSI device a particular slice of that drive). This ensures that each device exported by a given device driver has a unique minor number and name. That is, a driver manages a minor number-name space.
Each minor number, when combined with the major number of its parent driver, forms a dev.sub.-- t value that uniquely identifies each device. For example, the devices dev0, dev1, dev2 and dev3 managed by the driver sb at address 0 have respective dev.sub.-- t values of (32, 0), (32, 1), (32, 3) and (32, 3). The SpecFS 138s maintains a mapping of dev.sub.-- t values to their corresponding devices. As a result, all device open requests to the SpecFS identify the device to be opened using its unique dev.sub.-- t value.
The DevTree path to a device provides that device's physical name. For example, the physical name of the device dev0 is given by the string: /devices/iommu@addr0/sbus@addr1/esp1@addr3/sd@0:a, where sd@0:a refers to the device managed by the sd driver at address 0 whose minor.sub.-- name is a; i.e., the device dev0. The physical name identifies the special file 170 (shown in FIG. 2) (corresponding to an snode) that holds all of the information necessary to access the corresponding device. Among other things, the attributes of each special file 170 hold the dev.sub.-- t value associated with the corresponding device.
As mentioned above, a link.sub.-- generator 144 generates a device's logical name from the device's physical name according to a set of rules applicable to the devices managed by that link generator. For example, in the case of the device dev0 managed by the driver sd at address 0, a link generator for SCSI devices could generate the following logical name, /dev/dsk/c0t0d0s0, where c0 refers to the controller esp1@addr3, t0 refers to the target id the physical disk managed by the sd@0 driver, d0 refers to the sd@0 driver and s0 designates the slice with minor name a and minor number 0. The device dev0 associated with the sd@1 driver could be assigned the logical name, dev/dsk/c0t1d1s4, by the same link generator 144. Note that the two dev0 devices have logical names distinguished by differences in the target, disk and slice values. It is now described with reference to FIG. 4 how this infrastructure is presently employed in Solaris to enable an application to open a particular device residing on the computer 100.
Referring to FIG. 4, there is shown a flow diagram of operations performed in the memory 104 of the computer 100 by various operating system components in the course of opening a device as requested by an application 150. The memory 104 is divided into a user space 104U in which the applications 150 execute and a kernel space 104K in which the operating system components execute. This diagram shows with a set of labeled arrows the order in which the operations occur and the devices that are the originators or targets of each operation. Where applicable, dashed lines indicate an object to which a reference is being passed. Alongside the representation of the memory 104, each operation associated with a labeled arrow is defined. The operations are defined as messages, or function calls, where the message name is followed by the data to be operated on or being returned by the receiving entity. For example, the message (4-1), "open(logical.sub.-- name)," is the message issued by the application 150 asking the kernel 132 to open the device represented in the user space 104U by "logical.sub.-- name. In this particular example, the application is seeking to open the device dev2.
After receiving the open message (4-1), the kernel 132 issues the message (4-2), "get.sub.-- vnode(logical.sub.-- name)," to the file system 134. This message asks the file system 134 to return the vnode of the device dev2, which the kernel 132 needs to complete the open operation. In response, the file system 134 converts the logical name 166 to the corresponding physical name 168 using the logical name space 164. The file system 134 then locates the file designated by the physical name and determines the dev.sub.-- t value of the corresponding device from that file's attributes. Once it has acquired the dev.sub.-- t value, the file system 134 issues the message (4-3), "get.sub.-- vnode(dev.sub.-- t)," to the SpecFS 138s. This message asks the SpecFS 138s to return a reference to a vnode linked to the device dev2. Upon receiving the message (4-3) the SpecFS 138s creates the requested vnode 136 and an snode 136s, which links the vnode 136 to the device dev2, and returns the reference to the vnode 136 (4-4) to the file system 134. The file system 134 then returns the vnode reference to the kernel (4-5).
Once it has the vnode reference, the kernel 132 issues a request (4-6) to the SpecFS 138s to open the device dev2 associated with the vnode 136. The SpecFS 138s attempts to satisfy this request by issuing an open command (4-7) to driver 2, which the SpecFS knows manages the device dev2. If driver 2 is able to open the device dev2, it returns an open.sub.-- status message (4-8) indicating that the open operation was successful. Otherwise, driver 2 returns a failure indication in the same message (4-8). The SpecFS 138s then returns a similar status message (4-9) directly to the kernel 132. Assuming that "success" was returned in message (4-9), the kernel 132 returns a file descriptor to the application 150 that is a user space representation of the vnode 136 linked to the device dev2 (4-10). The application 150, once in possession of the file descriptor, can access the device dev2 via the kernel 132 and the file system 134 using file system operations. For example, the application 150 performs inputs data from the device dev2 by issuing read requests directed to the returned file descriptor. These file system commands are then transformed into actual device commands by the SpecFS 136s and the vnode and snode objects 136, 136s that manage the device dev2.
Consequently, Solaris enables users of a computer system 100 to access devices on that system 100 with relative ease. However, the methods employed by Solaris do not permit users to transparently access devices across computers, even when the different computers are configured as part of a cluster. That is, an application running on a first computer cannot, using Solaris, transparently open a device on a second computer.
The reason that the current version of Solaris cannot provide transparent device access in the multi-computer situation has to do with the way the dev.sub.-- t and minor numbers are currently assigned when devices are attached. Referring again to FIG. 3, each time a device is attached to the computer 100 the device's associated driver assigns that device a minor number that is unique within the set of devices controlled by that driver and therefore can be mapped to a unique dev.sub.-- t value for the computer 100 when combined with the driver's major number. However, if the same devices and driver were provided on a second computer, the driver and devices would be assigned a similar, if not identical, set of major and minor numbers and dev.sub.-- t values. For example, if both computers had a SCSI driver sd (major num=32) and four SCSI device instances managed by the SCSI driver sd, each driver sd would allocate the same set of minor numbers to their local set of SCSI devices (e.g., both sets would have minor numbers between 0 and 3). Consequently, keeping in mind that a device is accessed according to its dev.sub.-- t value, if a first node application wanted to open a SCSI disk on the second node, that application would not be able to unambiguously identify the SCSI disk to the SpecFS on either computer system.
Therefore, there is a need for a file-based device access system that enables applications, wherever they are executing, to transparently access devices resident on any node of a computer cluster.